SMA-LSTM Soft Sensor Based on Bayesian Optimization in Irregular Sampled Time Sequence of Industrial Processes

In modern industrial processes, soft sensor technology is used to predict quality parameters that are difficult to measure directly. Due to the complex dynamic relationships inherent in industrial process data series, time series models are often employed for predicting these quality parameters. However, most existing time series prediction models can only handle regular sampled data. Unfortunately, this condition dose not always hold true in real industrial environments. To address this issue, a sampling interval multi-attention LSTM (SMA-LSTM) model is proposed in this paper to handle with potentially irregular sampling conditions. Specifically, a set of attention mechanisms is applied to assess the impact of the sampling interval on the three gating signals. Interval information between sampling points is introduced into the network to adaptively adjust the model's estimation of the dynamic relationships between samples. Additionally, Bayesian optimization is employed to tune the hyperparameters of the network, making it easier to find reasonable parameters in different industrial processes. Experimental results demonstrate that the proposed SMA-LSTM is more effective than other time series prediction algorithms.